Per configuration

Report 21/06/2021

Questions

• Once noisy data is "cleaned" in first iteration then only use minimal renoising ($0\% \to 10\%$) for better training, i.e. simulates noisy input cable but then perfect "iterative" system. Not much learning from second iteration

Training structure

Loss

• Huber loss with $\delta =1$

Metrics

• Huber loss
• MNIST classifier (2 layers 32, 10 dense)

Noise visualization

Noise level mean comparissons

Huber loss

Classifier Accuracy

Detailed results per experiment

Noise with 0.0% to 0.1%

Noise with 0.2% to 0.3%

Noise with 0.4% to 0.5%

Noise with 0.6% to 0.7%

Noise with 0.8% to 0.9%